P1: GCR

Mathematical Geology [mg] PP344-matg-364915 January 16, 2002 14:59 Style file version June 30, 1999

Mathematical Geology, Vol. 34, No. 1, January 2002 (°

C 2002)

Conditional Simulation of Complex Geological

Structures Using Multiple-Point Statistics1
Sebastien Strebelle2

In many earth sciences applications, the geological objects or structures to be reproduced are curvil-
inear, e.g., sand channels in a clastic reservoir. Their modeling requires multiple-point statistics involv-
ing jointly three or more points at a time, much beyond the traditional two-point variogram statistics.
Actual data from the field being modeled, particularly if it is subsurface, are rarely enough to allow
inference of such multiple-point statistics. The approach proposed in this paper consists of borrowing
the required multiple-point statistics from training images depicting the expected patterns of geological
heterogeneities. Several training images can be used, reflecting different scales of variability and styles
of heterogeneities. The multiple-point statistics inferred from these training image(s) are exported to
the geostatistical numerical model where they are anchored to the actual data, both hard and soft,
in a sequential simulation mode. The algorithm and code developed are tested for the simulation of
a fluvial hydrocarbon reservoir with meandering channels. The methodology proposed appears to
be simple (multiple-point statistics are scanned directly from training images), general (any type of
random geometry can be considered), and fast enough to handle large 3D simulation grids.

KEY WORDS: geostatistics, stochastic simulation, training image, random geometry.

INTRODUCTION

This section briefly recalls the limitations of traditional simulation approaches

based on two-point statistics, and reviews present approaches which account for
multiple-point (mp) information. Other introductions to multiple-point (geo)statis-
tics can be found in Farmer (1988), Deutsch (1992), Journel (1997), and Srivastava
(1992, 1995).
Curvilinear geometries, such as sinuous channels in a fluvial reservoir or
incised valleys over a topography, cannot be modeled using only traditional two-
point statistics such as a variogram. Reproduction of such random geometries
calls for the parametrization of specific shapes or the consideration of the joint
categorical variability at three or more points at a time. This is the reason why

1Received30 March 2000; accepted 2 January 2000.

2Department of Geological and Environmental Sciences, Stanford University, Stanford, California
94305-2115; e-mail: strebell@pangea.stanford.edu

1
0882-8121/02/0100-0001/1 °
C 2002 International Association for Mathematical Geology
P1: GCR
Mathematical Geology [mg] PP344-matg-364915 January 16, 2002 14:59 Style file version June 30, 1999

2 Strebelle

specific geometries are poorly reproduced by traditional pixel-based algorithms,

such as indicator or Gaussian truncated simulation techniques which succeed at
reproducing only proportions and two-point (cross)variograms.
The most straightforward way to reproduce crisp geometries is to parametrize
their shapes, e.g. curvilinear “fettucinis” to model meandering channels, then ran-
domize these shape parameters and use Boolean object-based algorithms for drop-
ping these random geometries over the volume to be simulated (Bridge and Leeder,
1979; Haldorsen and Damsleth, 1990; Omre, 1991). There are two limitations to
such object-based approach: (1) each class of objects requires its own specific
parametrization and not all geomorphological heterogeneities can be summarized
by a few geometric parameters, (2) the conditioning to local data of these random
objects can be difficult, particularly if the data are dense with regard to the average
object size.
Coming back to the pixel-based alternative, Xu (1996) addresses the chal-
lenge of curvilinearity by presimulating local directions and ratios of geometric
anisotropy. The local kriging systems are then adapted to these locally varying
anisotropy characteristics, resulting in the simulation of changing anisotropy direc-
tions rather than curvilinear features. The technique being limited to reproduction
of two-point statistics fails to yield crisp continuous and curvilinear geometries.
Also Xu does not address the theoretically challenging problem of joint simula-
tion of a 3D stochastic direction defined by up to 3 related angles (azimuth, dip,
and rake).
Some of the pixel-based methods which use directly mp information to simu-
late curvilinear structures are

• Simulated annealing, where mp statistics inferred from a training image

are incorporated in the objective function (Deutsch, 1992; Farmer, 1988).
However, the number of components in the objective function increases
with the number of mp statistics imposed, consequently CPU and RAM
demand quickly increases and convergence becomes problematic.
• Markov Chain Monte Carlo (MCMC) simulation calls for a prior speci-
fication of either the mp probability distribution of the variable to be simu-
lated, or some ratio of conditional probability values given a mp data
event (Tjelmeland, 1996). An analytical definition of such mp probability
distributions typically requires severe model approximations such as data
screening, conditional independence and/or arbitrary Gaussian distribu-
tions with their congenial properties. In this regard, the MCMC implemen-
tation of Caers and Journel (1998) is notable in that it is free of most above
model assumptions because all required probability values are modeled
from training images.
• Iterative simulation algorithms based on a form of Gibbs sampler were
proposed by Srivastava (1992, 1995). Similar to the MCMC algorithms,
P1: GCR
Mathematical Geology [mg] PP344-matg-364915 January 16, 2002 14:59 Style file version June 30, 1999

Simulation of Complex Geological Structures Using Multiple-Point Statistics 3

they consist of iterative perturbations of a prior image one pixel at a time,

the perturbation ensuring that the mp statistics of a template centered on
the pixel being perturbed approximate the statistics inferred from a training
image.
All previous algorithms are iterative: the numerical model being built is perturbed
over multiple visits of each of its pixels, and they rely on convergence criteria.
If convergence is sometimes ensured by the properties of the Markov chain used,
the rate of convergence is rarely known a priori and somewhat arbitrary stopping
criteria must be implemented.
The paper by Guardiano and Srivastava (1993) represents a milestone in that
it suggests a direct (noniterative) algorithm for imposing mp statistics into stochas-
tic simulation. This algorithm consists of a sequential indicator simulation where
all required conditional probabilities are identified to corresponding proportions
read from training images. Guardiano and Srivastava’s idea is remarkable for its
extreme simplicity: no prior modeling of mp statistics or variogram is required, nor
any kriging to derive the conditional probabilities; these probabilities are obtained
directly by scanning the training image(s). However, although this algorithm is
noniterative and hence does not suffer from convergence considerations, the cor-
responding original code was extremely CPU demanding: the full training image
had to be scanned anew at each unsampled node to infer the node-specific con-
ditional probability distribution. The algorithm proposed in this paper is but an
extension of the seminal work of Guardiano and Srivastava.
Experience with stochastic simulation of categorical variables (e.g. facies
types or classes of a continuous variable) points towards developing new algo-
rithms which would combine the flexibility and easy data-conditioning of pixel-
based algorithms with the ability to reproduce “shapes” of object-based algorithms,
without being too CPU and RAM demanding. Ideally such an algorithm should be
pixel-based, include mp statistics allowing shape reproduction, be fast hence non-
iterative, and be general in that a new program need not be written to accommodate
any new random geometry.

Terminology

Consider an attribute S taking K possible states {sk , k = 1, . . . , K }. S can

be a categorical variable, or a continuous variable with its interval of variability
discretized into K classes by (K − 1) threshold values. A data event dn of size n
centered at a location u to be simulated is constituted by
• a data geometry defined by the n vectors {hα , α = 1, . . . , n}
• the n data values s(u + hα ) = s(uα ), α = 1, . . . , n

The value at the center of that data template is the unknown value s(u).
P1: GCR
Mathematical Geology [mg] PP344-matg-364915 January 16, 2002 14:59 Style file version June 30, 1999

4 Strebelle

A data template τn comprises only the previous data geometry. A subtemplate

of τn is a template constituted by any subset n 0 of vectors of τn , with n 0 ≤ n. The
data event dn is said to be “associated” with the geometric template τn .

Preliminary Remark

In a stochastic mode, the K possible outcomes of the random variable S(u)

are characterized by their conditional probability distribution function (cpdf) de-
noted as:

Prob{S(u) = sk | dn } = f (u; k | dn ), k = 1, . . . , K (1)

Knowledge of that cpdf for any data event dn suffices to generate by Monte Carlo
simulation realizations of the random variable S(u). Note, however, that such
knowledge is a tall proposition, indeed, if each of the n data variables can take K
different values, the total number of different data events for a given data geometry
is K n ; e.g. K = 10 decile classes and n = 10 lead to K n = 1010 , a huge number!
In the practice of sequential simulation, the previous cpdf is estimated at each
node u by some form of kriging, either kriging of the normal score transform of
the continuous variable S(u), or kriging of each of the K class indicators of the
categorical variable S(u) (Deutsch and Journel, 1998, p. 175).
An alternative to kriging would be to borrow the previous cpdf from a training
image T by scanning it for replicates of the data event dn . Suppose that J such
replicates are found, the histogram of the J central values sT (u j ), j = 1, . . . , J ,
can be used as a proxy for the cpdf f (u; k | dn ). The problem is that no finite
training image would be large enough to provide enough replicates for each of the
K n possible outcomes of the data event dn , if K n = 1010 . These are two avenues
of solutions around this inference problem:
1. The traditional modeling route whereby the cpdf f (u; k | dn ) is modeled
by some, preferably mp function of the n data values s(uα ). Instead of
some form of kriging, Caers and Journel (1998) utilize a neural network
to fit a mp parametric cpdf to the few (much lesser than K n !) experimental
proportions found over the training image.
2. Drastically reduce the total number K n by making K small, down to
K ≤ 4. Then only those proportions corresponding to the data events dn
actually found over the training image are utilized directly as cpdf values
without any prior modeling.
The last alternative pioneered by Guardiano and Srivastava (1993) is that retained
hereafter.
P1: GCR
Mathematical Geology [mg] PP344-matg-364915 January 16, 2002 14:59 Style file version June 30, 1999

Simulation of Complex Geological Structures Using Multiple-Point Statistics 5

A SINGLE NORMAL EQUATION

The key to any sequential simulation algorithm is the cpdf (1). In traditional
two-point algorithms, the conditioning is considered one datum at a time through
some measure of the two-point correlation between S(u) and S(uα ), for example
an indicator covariance model. Instead, we suggest considering jointly the n data
of the conditioning data event dn , which requires a (n + 1)-point covariance to
measure the dependence of S(u) on the data event dn .
Denoted by Ak the binary (indicator) random variable associated to the oc-
currence of state sk at location u:

½
1 if S(u) = sk
Ak =
0 if not

Similarly, let D be the binary random variable associated to the occurrence of the
data event dn constituted by the n conditioning data S(uα ) = skα , α = 1, . . . , n,
considered jointly:

½
1 if S(uα ) = skα , ∀α = 1, . . . , n
D=
0 if not

If the (n + 1)-point statistics relevant to Ak and its data event D are available,
then the exact conditional probability is given by the simple kriging expression
(Journel, 1993):

Prob{Ak = 1 | D = 1} = E{Ak } + λ[1 − E{D}] (2)

where D = 1 is the observed data event, E{D} = Prob{D = 1} is the probability

for the conditioning data event to occur, E{Ak } = Prob{S(u) = sk } is the prior
probability for the (unknown) state at u to be sk ; this probability is “prior” to
knowledge of the data event D = 1.
The single extended normal (kriging) equation providing the single weight λ
is written as

λ Var{D} = Cov{Ak , D} (3)

where Cov{Ak , D} = E{Ak D} − E{Ak }E{D} is a (n + 1)-point statistics.

P1: GCR
Mathematical Geology [mg] PP344-matg-364915 January 16, 2002 14:59 Style file version June 30, 1999

6 Strebelle

Then, as per the kriging Equation (3):

E{Ak D} − E{Ak }E{D}

λ= , leading to the solution:
E{D} (1 − E{D})
E{Ak D} − E{Ak }E{D}
Prob{Ak = 1 | D = 1} = E{Ak } +
E{D}
E{Ak D} Prob{Ak = 1, D = 1}
= = (4)
E{D} Prob{D = 1}

That exact solution identifies the definition of the conditional probability, as given
by Bayes’ relation.

Scanning the Training Image

The exact solution (4) calls for (n + 1)-point statistics much beyond the tra-
ditional two-point variogram or convariance model. There is usually no hope to
be able to infer such mp statistics from actual sample data, hence the idea to bor-
row them by scanning one or several training images under a prior decision of
stationarity (export license):
• The denominator Prob{S(uα ) = skα , α = 1, . . . , n} of expression (4) can
be inferred by counting the number c(dn ) of replicates of the conditioning
data event dn = {S(uα ) = skα , α = 1, . . . , n} in the training image(s). A
replicate should have same geometric configuration and same data values.
• The numerator Prob{S(u) = sk and S(uα ) = skα , α = 1, . . . , n} is obtained
by counting the number ck (dn ) of replicates, among the c previous ones,
associated to a central value S(u) equal to sk .
The required conditional probability is then identified to the training proportion
ck (dn )/c(dn ):

ck (dn )
p(u; sk | (n)) = Prob{Ak = 1 | D = 1} = Prob{S(u) = sk | (n)} ' (5)
c(dn )

The limit case of a single extended normal (kriging) Equation (3) is absolutely
straightforward in that it reduces to the very definition of a conditional probability,
which is obtained by scanning the training image with the conditioning data tem-
plate. The intermediary step of modeling permissible functions to sample statistics,
e.g. a covariance model, is completely shortcut. The conditional probability (4) is,
by definition, permissible, i.e. it does not suffer any of the order relation problems
associated to determining an estimate of it through a kriging limited to two-point
statistics. Note, however that expression (4) is “exact” only up to the decision of
P1: GCR
Mathematical Geology [mg] PP344-matg-364915 January 16, 2002 14:59 Style file version June 30, 1999

Simulation of Complex Geological Structures Using Multiple-Point Statistics 7

stationarity which allows exporting mp statistics from the training image to the
actual phenomenon under study.
An important aspect of the extended normal equation approach relates to
which mp statistics should be retained to condition the simulation of any partic-
ular node u. The larger the data neighborhood, the larger the size n of the data
event dn , the more specific this data event, hence the fewer replicates of it will be
found over the training image(s) for inference of the corresponding conditional
probability p(u; sk | (n)). Also, if dn is too specific, the probability distribution
{ p(u; sk | (n)), k = 1, . . . , K } may be too specific to the training image(s) retained,
hence exporting it to the model may be questionable: we may be exporting idiosyn-
cratic patterns of the training image instead of its essence. The approach proposed
hereafter consists of always inferring the full multiple (n + 1)-point statistics from
the training image, reducing progressively the size n until the data event dn retained
is found “often enough” in the training image(s). This implies that the training im-
age has a repetitive character, i.e. the geometrical shapes to be reproduced are
repeated several times, say 10–20 times, in the training image. The minimal num-
ber cmin of replicates of the conditioning data event dn to be found in the training
image for the corresponding cpdf to be retained is then related to the repetitive-
ness of the training image, hence cmin may be, e.g., 10 or 20. Consequently the
stationarity decision which allows exporting mp statistics from the training image
refers to the essential (i.e. commonly found) features of the training image.
A critical aspect of this approach is its total dependence on the training im-
age(s) used. The structures are borrowed directly from that training image, without
any modeling or filtering, and anchored to the actual data. The training image is
used here as a replacement for the implicit mp distribution (spatial law) underlying
any mapping algorithm, whether deterministic as simple as hand-contouring, or
stochastic. One could argue that a training image has the advantage of making
fully explicit the structural information being exported to the mapping of the ac-
tual phenomenon. That prior information can be easily evaluated, then accepted
or rejected, whereas it is much more difficult to evaluate the appropriateness of,
say, a Gaussian or Gaussian-related random function model.
Training images depict the patterns of geological heterogeneities deemed rele-
vant to the application under study. They need not carry any locally accurate infor-
mation on the actual phenomenon; they merely reflect a prior geological/structural
concept. Thus, a training image can be an unconditional realization generated by an
object-based algorithm, or a simulated realization of an analogous field, or simply
a geologist’s sketch processed with CAD algorithms and properly digitized.

THE SNESIM ALGORITHM

We chose to name the simulation algorithm presented hereafter snesim to in-

sist that it involves only one single normal equation (sne), which leads to the very
P1: GCR
Mathematical Geology [mg] PP344-matg-364915 January 16, 2002 14:59 Style file version June 30, 1999

8 Strebelle

expression (4) of a conditional probability. The snesim algorithm has been devel-
oped to simulate categorical attributes, e.g. geological facies, but can be extended to
simulate continuous attributes discretized into a small finite number K of classes.
The snesim algorithm is based on the sequential simulation paradigm whereby
each simulated value becomes a hard datum value conditioning the simulation of
nodal values visited later in the sequence (Goovaerts, 1997, p. 376). Because
the local conditioning data event includes previously simulated nodes, and nodes
are visited along a random path, the geometry of that data event changes from one
node to the other. Guardiano and Srivastava (1993) proposed scanning the complete
training image anew at each unsampled node to infer the conditional probability
distribution specific to the data informing that node. Such repetitive scanning can
be very CPU demanding, especially when considering a large training image or
when generating a large number of realizations each with many nodes.
The algorithm implementation proposed in this paper is much less CPU de-
manding without being too memory (RAM) demanding. This new implementation
is based on the two following properties:

Property 1. Given a template τn of n data variables, the number of cpdf’s

associated with τn (pdfs conditional to data events dn associated with τn ) that
can be actually inferred from the training image is related to the training image
dimensions, hence is generally much smaller than the total number K n of cpdfs
associated with τn .

Property 2. The probability distribution conditional to a data event dn 0 , asso-

ciated with a subtemplate τn 0 of τn (n 0 ≤ n) can be retrieved from the probability
distributions conditional to the data events dn associated with τn and for which
dn 0 is subset.

Let dn 0 be a data event associated with a subtemplate τn 0 of τn (n 0 ≤ n). The number

c(dn 0 ) of replicates of dn 0 is equal to the sum of the replicates of all data events dn
associated with τn and for which dn 0 is a subset (Strebelle, 2000):
X
c(dn 0 ) = c(dn )
dn ass. with τn
dn 0 ⊂dn

Similarly for the number ck (dn 0 ) of dn -replicates with a central value S(u) equal
to sk :
X
ck (dn 0 ) = ck (dn )
dn ass. with τn
dn 0 ⊂dn

Knowledge of ck (dn 0 ) and c(dn 0 ) allows then estimating the probability distribution
conditional to dn 0 , using relation (5).
P1: GCR
Mathematical Geology [mg] PP344-matg-364915 January 16, 2002 14:59 Style file version June 30, 1999

Simulation of Complex Geological Structures Using Multiple-Point Statistics 9

Denote by W (u) the data search neighborhood centered on location u.

Consider the data template τn constituted by the n vectors {hα , α = 1, . . . , n}
defined such that the n locations u + hα , α = 1, . . . , n correspond to the n grid
nodes present within W (u). The snesim algorithm proceeds in two steps:

1. First, store in a dynamic data structure, called search tree (Roberts, 1998,
p. 537–592), only those cpdfs associated with τn that can be actually
inferred from the training image. More precisely, store in a search tree the
numbers of occurrences of data events and central values (ck (dn )) actually
found over the training image, and from which the training proportions
(5) can be calculated. Because of property 1, the amount of RAM required
by the search tree is not too large, if a data template τn with a reasonable
number n of nodes, say, less than 100 nodes, is retained. The construction
of that search tree requires scanning the training image one single time
prior to the image simulation, hence it is very fast (Strebelle, 2000).
2. Next, perform simulation by visiting each grid node one single time along
a random path. At each node u to be simulated, the conditioning data are
searched in W (u), hence the local conditioning data event is necessarily
associated with a subtemplate of τn . According to property 2, local cpdf can
be retrieved from the search tree. The training image need not be scanned
anew at each unsampled node, which renders the snesim algorithm much
faster than Guardiano and Srivastava’s original implementation.

The main steps of the snesim simulation algorithm are now presented in detail:

1. Scan the training image(s) to construct the search tree. Only those data
events which actually occur over the training image are stored in the search
tree. A maximum data search template is defined to limit the geometric
extent of those data events.
2. Assign the original sample data to the closest grid nodes. Define a random
path visiting once and only once all unsampled nodes.
3. At each unsampled location u, retain the conditioning data actually present
within the maximum search template used to construct the search tree. Let
n 0 be the number of those conditioning data, and dn 0 the corresponding data
event. Retrieve from the search tree the proportions of type (5) correspond-
ing to the data event dn 0 . To ensure that these proportions are significant, if
the total number c(dn 0 ) of training dn 0 -replicates is less than an input mini-
mum value cmin , the most distant conditioning datum is dropped, reducing
the number of conditioning data to (n 0 − 1); proportions conditioned to
this lesser data event dn 0 −1 are retrieved again from the search tree, and so
on . . . If the number of data drops to n 0 = 1, and c(dn 0 ) is still lower than
cmin , the conditional probability p(u; sk | (n 0 )) is replaced by the marginal
probability pk .
P1: GCR
Mathematical Geology [mg] PP344-matg-364915 January 16, 2002 14:59 Style file version June 30, 1999

10 Strebelle

4. Draw a simulated s-value for node u from the cpdf read from the search
tree. That simulated value is then added to the s-data to be used for con-
ditioning the simulation at all subsequent nodes.
5. Move to next node along the random path and repeat steps 3 and 4.
6. Loop until all grid nodes are simulated. One stochastic image has been
generated. Reiterate the entire process from step 2 with a different random
path to generate another realization.

To ensure exact reproduction of the hard data at their locations, all original
sample data are relocated to the nearest simulation grid node and their values are
frozen. The small scale spatial continuity of the training image is passed to the
simulated realizations through the training proportions, hence as the node being
simulated gets closer to a hard datum location the conditional variance decreases
as it does on the training image.

Multiple Grid Implementation

The maximum data search template retained should not be taken too small,
otherwise large scale structures of the training image would not be reproduced.
On the other hand, a search template including too many grid nodes would lead to
storing a large number of cpdfs in the search tree, increasing CPU cost and memory
demand. One solution to capture large scale structures while considering a data
search template with a reasonably small number of grid nodes is provided by the
multiple grid concept initially proposed by Gómez-Hernández (1991) and further
developed by Tran (1994). The multiple grid approach implemented in snesim
consists of simulating a number G of increasingly finer grids. The gth (1 ≤ g ≤ G)
grid is constituted by each (2g−1 )th node of the final simulation grid (g = 1). The
data search template adopted for the various nested simulation grids need not
have the same geometric configuration. The larger search templates of the coarser
simulation grids allow capturing the large scale structures of the training image.
One search tree needs to be constructed per simulation grid, possibly using a
different training image reflecting the heterogeneities specific to that scale. When
the gth grid simulation is completed, its simulated values are frozen as data values
to be used for conditioning on the next finer simulation grid.

Reproduction of the Marginal Distribution

In snesim, no explicit constraint ensures reproduction of the sample his-

togram, or any other target marginal distribution { pk , k = 1, . . . , K }. Simulated
realizations may display global proportions significantly different from the origi-
nal sample distribution, particularly if probability values are borrowed from a
P1: GCR
Mathematical Geology [mg] PP344-matg-364915 January 16, 2002 14:59 Style file version June 30, 1999

Simulation of Complex Geological Structures Using Multiple-Point Statistics 11

training image having a histogram significantly different from the target marginal
distribution.
Programs, such as the GSLIB program trans (Deutsch and Journel, 1998,
p. 227), allow postprocessing of any training image or simulated realization to
approximate any target histogram, while honoring the original hard data values at
their locations and without affecting the values rank orders. However, if the target
histogram is very different from the original image histogram, both variograms
and mp statistics may be changed significantly by this transform.
The first recommendation is to generate training images with global pro-
portions reasonably similar to the target proportions one wish to impose to the
final model. Next a servosystem, described in detail in Strebelle (2000), has been
implemented which modifies gradually the proportions (5) so that the simulation
algorithm always remains close to the global target proportions { pk , k = 1, . . . , K }
as it progresses from one node to another.

Integration of Secondary Data

Conditioning could also include soft information, e.g. seismic data in reser-
voir applications. Bayes’ relation (4) can be extended to evaluate the probabil-
ity distribution of a variable S(u) conditioned to both the nearest n hard data
s(u + hα ) = skα , α = 1, . . . , n, and the collocated soft datum y(u) = y:

Prob{Ak = 1, D = 1, y}
Prob{Ak = 1 | D = 1, y} = (6)
Prob{D = 1, y}

Starting from a training image of the primary variable, the corresponding training
image of the soft data variable can be forward simulated. Both hard and soft
training images are considered as a single vectorial training image T. Under a
prior decision of stationarity (export license), this vectorial training image can be
scanned to evaluate expression (6). In practice, this scanning requires reducing
the y-conditioning to only a few locations of y-data values, for example the sole
collocated value y(u) as in expression (6), and a few classes of y-values, for
example the four quartile classes of the y-sample histogram.
Note that D includes the previously simulated nodes; hence a hard data train-
ing image is required even if no original sample hard data are available.
Extension of the algorithm to conditioning to more than one soft y-datum
is absolutely straightforward. Increasing the number n Y of conditioning soft data,
however, entails increasing the size of the extended conditioning data event dn . If
each of the n hard data variables can take K different values and each of the n Y
soft data variables can take L different values, the total number of different data
events for a given data geometry is K n L n Y , which increases dramatically with n Y .
The number c(dn+n Y ) of training replicates of the extended conditioning data event
P1: GCR
Mathematical Geology [mg] PP344-matg-364915 January 16, 2002 14:59 Style file version June 30, 1999

12 Strebelle

decreases dramatically with n Y , which renders the inference of the local probabil-
ity distribution more difficult. The search tree relates only to those extended data
events which actually occur over the vectorial training image.
If n Y has to be large, e.g. to allow capturing patterns or mp statistics displayed
by the soft data, a different algorithm is proposed in Strebelle (2000). This algo-
rithm amounts to combine the snesim-derived probability Prob{Ak = 1 | D = 1}
with the probability Prob{Ak = 1 | y}, the latter being possibly derived by neural
net modeling of a pair of training images (facies vs. soft data).

SIMULATION OF A FLUVIAL RESERVOIR

The snesim algorithm was tested on the simulation of a horizontal 2D section

of a fluvial reservoir. Fluvial reservoirs are characterized by the presence of sinuous
sand-filled channels within a background of mudstone (Facies 1). For this example,
we consider two types of sand: border sands (Facies 2) constituted by levies and
crevasses splays which are intermediate in reservoir quality, and channel sands
(Facies 3) which correspond to the best reservoir rock. A realistic modeling of the
curvilinear sand channel patterns is critical for reliable connectivity assessment
and flow simulation of such reservoir.
The object-based program fluvsim (Deutsch and Wang, 1996) was used to
generate the reference “true” image of size 100 ∗ 100 = 10000 pixels shown in
Figure 1(A). Thirty-two sample data were collected at random over the true image
with facies proportions close to the true global proportions (Fig. 1(B)).
Two training images which could have been hand drawn by a geologist, then
digitized, were considered.
• A first large scale training image, comprising 100 ∗ 100 = 10000 pixels,
depicts the deemed geometry of sand channels versus their background
constituted by Facies 1 and 2 pooled together (Fig. 1(C)). This image
carries information about the mean width, the major direction, and the
sinuosity of the channels and their grouping in space.
• A second small scale training image, comprising only 35 ∗ 35 = 1225
pixels, carries the prior information about the mean width of the levies
and crevasses, and their locations relative to the channels (Fig. 1(D)).
Simulation proceeds in two nested sequences:
• First, Facies 3 (channel sands) was simulated versus Facies 1 and 2 pooled
together, conditional to the 32 sample channel indicator data of Figure 1(B),
using the large scale training image. The multiple grid option was used to
capture the large scale channel continuity: four increasingly finer grids
were considered.
P1: GCR
Mathematical Geology [mg] PP344-matg-364915 January 16, 2002 14:59 Style file version June 30, 1999

Simulation of Complex Geological Structures Using Multiple-Point Statistics 13

Figure 1. Simulation of a horizontal 2D section of a fluvial reservoir. (A) True image, (B) 32
sample data, (C) large scale training image, (D) small scale training image.

Ten realizations were generated using snesim. Their generation took

2.4 CPU second per realization with a DEC 600 MHz desktop computer un-
der Unix. About 17 nearby data were retained in average for the simulation
of any single node.
The first snesim realization is displayed in Figure 2(A). To provide
a yardstick for comparison, Figure 2(B) shows a realization conditioned
to the same 32 sample data, generated by the GSLIB sequential indicator
simulation program sisim (Deutsch and Journel, 1998, p. 175), hence using
a variogram model accounting only for two-point correlation. In contrast
to sisim, snesim allows reproducing reasonably well the sinuous channel
patterns displayed by the training image, although with some deficit of
large scale continuity: the simulated channels do not all cross the image
from one side to the other. This is a problem of ergodicity which can be
alleviated by using a much larger training image, at least twice larger than
the maximum distance of continuity of the channels.
The servosystem described in Strebelle (2000) allows one to repro-
duce reasonably well the sample channel proportion p̂ = 28.1%: the sand
proportions of the ten realizations fluctuate between 27.3 and 28.3% with
a mean of 27.7%. Note that the difference between the training and the tar-
get sample channel proportions is greater than 4%. This indicates that the
P1: GCR
Mathematical Geology [mg] PP344-matg-364915 January 16, 2002 14:59 Style file version June 30, 1999

14 Strebelle

Figure 2. Simulation of Facies 3 (channel sand). (A) One snesim realization, (B) one sisim re-
alization, (C) multiple-point connectivity along EW (continuous line refers to true image, dashed
lines refer to the 10 snesim realizations), (D) EW channel sand indicator variograms standardized
to variance 1 (continuous line refers to true image, dashed lines refer to the 10 snesim realizations).

algorithm can handle training images with inaccurate global proportions

as long as they depict accurately the geometry of heterogeneities.
A static measure of connectivity of the sand channels can be obtained
through the mp connectivity function proposed by Journel and Alabert
(1989). This function is the proportion of connected strings of n channel
pixels plotted versus n. Figure 2(C) displays this connectivity function
calculated along the EW direction of channel orientation for the true image
and the 10 snesim realizations. All realizations reproduce reasonably well
the true connectivity of channels.
Although there is no variogram modeling nor explicit constraint for
variogram reproduction in snesim, the simulated channel sand indicator
variograms along the EW direction are close to the variogram of the refer-
ence image (Fig. 2(D)). The two-point statistics are implicitly included in
the mp statistics exported to the simulated image.
• Next Facies 1 and 2 were simulated within the areas previously simulated
as Facies 1 + 2 (nonchannel sand), conditional to the original Facies 1 and
2 hard data and the previously simulated Facies 3 nodal values, using the
small scale training image. When scanning that small scale training image,
data events for which the central value is Facies 3 are discarded. Hence,
P1: GCR
Mathematical Geology [mg] PP344-matg-364915 January 16, 2002 14:59 Style file version June 30, 1999

Simulation of Complex Geological Structures Using Multiple-Point Statistics 15

Figure 3. Simulation of Facies 1 and 2 (mudstone and border sand). Two snesim simulated
realizations.

the conditional probability for Facies 3 is zero, which ensures that Facies 3
cannot be simulated in the Facies 1 + 2 area.

The simulated realizations corresponding to the first two channel realizations

generated in the first step are shown in Figure 3. Because levies were treated
as small scale heterogeneities, their actual large scale continuity is understated.
Conversely, the crevasse splays being lumped together with the levies are simulated
more elongated than they are over the small scale training image. Note that the
resulting simulated Facies 2 (border sand) is correctly attached to the channels.

Sensitivity to the Training Image

To analyze the sensitivity of the snesim simulated realizations to the training

image, sand channels are now simulated using the two alternative large scale
training images displayed in Figure 4(A) and (C). The first alternative train-
ing image displays much thinner channels than the original training image of
Figure 1(C). The second alternative training image was simply obtained by a 90◦
rotation of the original training image. One snesim realization was generated using
each alternative training image (Fig. 4(B) and (D)). In both cases, the simulated
channels reproduce the shape of the corresponding training channels: they display
the smaller thickness of the first alternative training image, or the NS preferential
orientation of the second alternative training image.
Those results show the dependence of the simulated realizations on the train-
ing image. snesim anchors the mp structures borrowed from the training image to
the original sample data. The “essence” of the training image, constituted by the
geometrical patterns found often enough in that image and consistent with the sam-
ple data, is reflected by the simulated realizations. The training image determines
the main features of the simulated images, however less, so as the sample data
are more numerous. Indeed hard data are frozen at their locations; hence if they
are many, the structures they reveal, especially medium to large scale structures,
P1: GCR
Mathematical Geology [mg] PP344-matg-364915 January 16, 2002 14:59 Style file version June 30, 1999

16 Strebelle

Figure 4. Sensitivity of the snesim simulated realizations to the training image. (A and B) First
alternative large scale training image and corresponding simulated realization conditional to the 32
sample data of Figure 1B, (C and D) second alternative training image and corresponding simulated
realization conditional to the 32 sample data of Figure 1B, (E) set of the 200 sample data collected
at random from the reference image of Figure 1A; (F) simulated realization conditional to the 200
sample data using the second alternative training image.
P1: GCR
Mathematical Geology [mg] PP344-matg-364915 January 16, 2002 14:59 Style file version June 30, 1999

Simulation of Complex Geological Structures Using Multiple-Point Statistics 17

prevail over the structures read from the training image and conflicts may occur
resulting in poorer reproduction of certain structures of the training image.
To illustrate such possible conflict, 200 sample data were collected at random
over the true image of Figure 1(A) (Fig. 4(E)), and a simulated realization condi-
tional to those 200 sample data was generated using the second alternative training
image shown in Figure 4(C). Because the 200 sample data reflect the actual EW
continuity, the large scale NS-oriented training channel patterns are reproduced
poorly (Fig. 4(F)). Such poor reproduction of the structures of the training image
is indicative of either unreliable hard data or an inappropriate training image.
The sensitivity of the snesim simulated realizations to the size and the geo-
metric configuration of the data template used to construct the search trees is
analyzed in Strebelle (2000): the snesim algorithm appears robust with respect to
the data template retained, as long as it is not too small.

SIMULATION OF MULTIPLE COMPLEX PATTERNS

In contrast to object-based techniques, the algorithm proposed in this paper

is general: snesim allows simulating any type of structures, of any shape, at any
scale, as long as the training image has a repetitive character. A new program need
not be written to fit any particular parametric object type or geometry. To illustrate
this point, a second case study is proposed.
The reference “true” image and the training image, both of size 500 ∗ 500 =
250,000 pixels, display a mixture of three geometrically different patterns
(Fig. 5(A) and (C)):

• ellipses elongated in the EW direction (Facies 1);

• crescents in the diagonal direction (Facies 2);
• small crosses (Facies 3).

Background is Facies 4. Crosses do not overlap each other, nor any ellipse or
crescent. Crescents overlap ellipses.
A large number of sample data (400 or 0.16% of the true image) were collected
at random over the true image, with facies proportions close to the true global
proportions (Fig. 5(B)). Figure 5(D) displays the corresponding training image for
the ellipses (Facies 1) only. Simulation proceeds in three nested sequences:

• First, Facies 1 (ellipses) is simulated versus all other facies pooled together,
and conditional to the 400 Facies 1 sample indicator data of Figure 5(B).
The conditional probability distributions are inferred from the ellipse train-
ing image shown in Figure 5(D), itself reconstituted from the full training
image of Figure 5(C). Anisotropy of the ellipses is accounted for by con-
sidering an anisotropic search template elongated along EW to construct
the search tree.
P1: GCR
Mathematical Geology [mg] PP344-matg-364915 January 16, 2002 14:59 Style file version June 30, 1999

18 Strebelle

Figure 5. Simulation of multiple patterns. (A) True image, (B) location of the 400 sample data,
(C) training image, (D) reconstructed training image used to simulate Facies 1, (E and F) two
simulated realizations.

• Next, Facies 2 (crescents) is simulated versus all other facies pooled to-
gether. This simulation is independent of that of Facies 1, and is conditional
to the 400 Facies 2 sample indicator data of Figure 5(B). The conditional
probability distributions are inferred from the training image of Figure 5(C),
Facies 1, 3, and 4 being pooled together. To account for the anisotropy of
the crescents, an anisotropic search template elongated along N45E was
used to construct the search tree.
Because crescents overlap the ellipses in the training image of
Figure 5(C), the crescent realizations were superimposed on top of the
P1: GCR
Mathematical Geology [mg] PP344-matg-364915 January 16, 2002 14:59 Style file version June 30, 1999

Simulation of Complex Geological Structures Using Multiple-Point Statistics 19

ellipse realizations generated in the first step, resulting in a realization

displaying ellipses and crescents versus Facies 3 and 4 background.
• Finally, since Facies 3 (crosses) does not overlap ellipses or crescents,
it is simulated versus Facies 4 only within areas previously simulated as
Nonfacies 1 and 2, and conditional to the 400 Facies 3 sample indicator
data of Figure 5(B). Inference of the conditional probability distributions
is limited to the Facies 3 and 4 areas of the training image displayed in
Figure 5(C). An isotropic conditioning search template was retained to
construct the corresponding search tree.

The previous sequence of simulations was used to generate the two realiza-
tions of Figure 5(E) and (F). The three types of structures are reasonably reproduced
with facies proportions close to the target sample proportions of Figure 5(B). Al-
though simulation proceeds in three nested sequences, the same general algorithm,
namely snesim, has been used to simulate the three types of patterns displayed by
the training image of Figure 5(C).
Recall that the snesim simulation algorithm is exact, in that it honors all hard
data values at their exact locations, no matter the density or configuration of these
data. In presence of the dense data shown in Figure 5(B), this exactitude condition
would be difficult to fulfill with object-based algorithms.

PRELIMINARY CONCLUSIONS

A new algorithm (snesim) is proposed for sequential simulation of categorical

random fields. This algorithm utilizes local proportions read from training images.
In contrast to classical algorithms based only on the two-point variogram model,
snesim allows reproduction of complex multiple-point patterns, such as undulating
channels. Such multiple-point patterns include the traditional 2-point statistics
which are thus also reproduced even though they are not modeled explicitly. The
algorithm requires large amount of RAM but not exceeding what is presently (1999)
available on desktop computers, and is fast enough to consider 3D simulation grids.
In contrast to object-based techniques, the proposed approach is general: the
same algorithm allows simulation of any type of geological heterogeneity, of any
shape, at any scale. snesim does not need to be modified to fit any particular para-
metric object type or geometry. The reason for such generality is that the user need
not decide in advance which mp statistics or geometric parameters are essential and
should be reproduced in the simulated realizations. All statistics are directly pro-
vided by the training images which, therefore, should display the heterogeneities
deemed relevant to the phenomenon being modeled. The classical steps of vari-
ogram modeling and kriging or object parametrization are completely shortcut.
The simplicity of the snesim algorithm comes, however, from a greater re-
liance on the prior decision of stationarity: much more than a mere variogram is
P1: GCR
Mathematical Geology [mg] PP344-matg-364915 January 16, 2002 14:59 Style file version June 30, 1999

20 Strebelle

borrowed from the training image(s). As of now, the snesim algorithm borrows
all structures present in the training images without discrimination at the risk of
exporting details irrelevant to the actual phenomenon being modeled. The extent
of the export license (stationary decision) raises two important questions: which
mp statistics should be borrowed from the training image, and how can we limit
the patterns being borrowed to those specific statistics?
The first question calls for a better understanding of the “essence” of a training
image: could some objective criteria, such as impact on flow response in reservoir
applications, be proposed to identify those “essential” mp statistics to be exported?
The on-going work of Caers, Srinivasan, and Journel (1999) explores this avenue.
Or should we rely on a more subjective (visual) identification of those patterns to
be reproduced?
The second question could be addressed by a proper selection of the con-
ditioning data events. Restricting the mp statistics which can be borrowed from
the training image entails constraints on the geometry of the conditioning data
events. The present approach, which consists of dropping the furthest away datum
whenever the probability distribution cannot be inferred reliably from the training
image, does not allow such control on the geometry of the data events.

REFERENCES

Bridge, J. S., and Leeder, M. R., 1979, A simulation model of alluvial stratigraphy: Sedimentology,
v. 26, p. 617–644.
Caers, J., and Journel, A. G., 1998, Geostatistical quantification of geological information for a fluvial-
type North-Sea reservoir: SPE paper no. 56655.
Caers, J., Srinivasan, S., and Journel, A. G., 1999, Stochastic reservoir simulation using neural networks
trained on outcrop data: SPE paper no. 49026.
Deutsch, C. V., 1992, Annealing techniques applied to reservoir modeling and the integration of geo-
logical and engineering (well test) data: Unpublished doctoral dissertation, Stanford University,
306 p.
Deutsch, C. V., and Journel, A. G., 1998, GSLIB: Geostatistical software library and user’s guide,
2nd ed.: Oxford University Press, New York, 368 p.
Deutsch, C. V., and Wang, L., 1996, Hierarchical object-based stochastic modeling of fluvial reservoirs:
Math. Geol., v. 28, no. 7, p. 857–880.
Farmer, C. L., 1988, The generation of stochastic fields of reservoir parameters with specified geostatis-
tical distributions, in Edwards, S., and King, P. R., eds., Mathematics in oil production: Clarendon
Press, Oxford, p. 235–252.
Gómez-Hernández, J. J., 1991, A stochastic approach to the simulation of block conductivity fields
conditioned upon data measured at a smaller scale: Unpublished doctoral dissertation, Stanford
University, 351 p.
Goovaerts, P., 1997, Geostatistics for natural resources evaluation: Oxford University Press, New York,
483 p.
Guardiano, F., and Srivastava, R. M., 1993, Multivariate geostatistics: Beyond bivariate moments, in
Soares, A., ed., Geostatistics-Troia, Vol. 1: Kluwer Academic, Dordrecht, p. 133–144.
Haldorsen, H. H., and Damsleth, E., 1990, Stochastic modeling: J. Pet. Technol., v. 42, April,
p. 404–412.
P1: GCR
Mathematical Geology [mg] PP344-matg-364915 January 16, 2002 14:59 Style file version June 30, 1999

Simulation of Complex Geological Structures Using Multiple-Point Statistics 21

Journel, A. G., 1993, Geostatistics: Roadblocks and challenges, in Soares, A., ed., Geostatistics-Troia,
Vol. 1: Kluwer Academic, Dordrecht, p. 213–224.
Journel, A. G., 1997, Deterministic geostatistics: A new visit, in Baffi, E., and Shofield, N., eds.,
Geostatistics Wollongong, Vol. 1: Kluwer Academic, Dordrecht, p. 174–187.
Journel, A. G., and Alabert, F. G., 1989, Non-Gaussian data expansion in the Earth Sciences: Terra
Nova, v. 1, p. 123–134.
Omre, H., 1991, Stochastic models for reservoir characterization, in Kleppe, J., and Skjaeveland, S. M.,
eds., Recent advances in improved oil recovery methods for North Sea sandstone reservoirs:
Norwegian Petroleum Directorate, Stavanger, 14 p.
Roberts, E. S., 1998, Programming abstractions in C: A second course in computer science: Addison-
Wesley, Reading, MA, 819 p.
Srivastava, M., 1992, Iterative methods for spatial simulation: Stanford Center for Reservoir Forecast-
ing, Rep. No. 5, 24 p.
Srivastava, M., 1995, An overview of stochastic methods for reservoir characterization, in Yarus, J., and
Chambers, R., eds., Stochastic modeling and geostatistics: Principles, methods, and case studies,
Vol. 3: AAPG Computer Applications in Geology, Tulsa, p. 3–16.
Strebelle, S., 2000, Sequential simulation drawing structures from training images: Unpublished doc-
toral dissertation, Stanford University, 200 p.
Tjelmeland, H., 1996, Stochastic models in reservoir characterization and Markov random fields
for compact objects: Unpublished doctoral dissertation, Norwegian University of Science and
Technology.
Tran, T. T., 1994, Improving variogram reproduction on dense simulation grids: Comput. Geosci., v. 20,
no. 7, p. 1161–1168.
Xu, W., 1996, Conditional curvilinear stochastic simulation using pixel-based algorithms: Math. Geol.,
v. 28, no. 7, p. 937–949.